HOW TO HANDLE UNKNOWN INPUTS

Abstract

This paper presents an empirical comparison among five different methods of handling missing inputs. Fiteen different problems are used for the comparison, and the performance of the methods is obtained for several percentages 0%, 5%, 10%, 20%, 30% and 40% of missing inputs in the problem. There is one method, which has the best performance in all the problems. The method is based on a gereralization of Backpropagation to interval arithmetic, a training including missing inputs in the training set, and a codification during training of every missing input by an interval which lower and uppper limits include the range of variation of the input.